AI Prompts for Startups: The Early-Stage Founder's Week (2026)

TL;DR: The AI prompts that help an early-stage startup assume no data, no team and no brand guide. Use AI for structure and drafting across discovery, positioning, pricing, landing copy, outreach, investor updates and hiring. Do not use it to size your market, check competitor facts, or tell you whether your idea is good.

What AI prompts actually work for an early-stage startup?

The ones that hand the model your evidence and ask for structure. Not the ones that ask the model for facts you do not have.

That single distinction sorts almost every founder prompt into useful or dangerous. "Here are six interview transcripts, find the pattern" is useful. "What is the market size for X" is dangerous, and it is dangerous in a way that looks exactly like being useful.

At the early stage you are missing the three things most prompt libraries quietly assume. You have no historical data, so nothing can be analysed. You have no team, so there is no second reader to catch a confident error before it reaches an investor. You have no brand guide, so an instruction like "match our tone of voice" resolves to nothing and the model falls back to its house style. A prompt written for a Series B marketing team fails on all three counts.

There is a second difference, and it is the one people miss. An operator is optimising a process that already works, so the model's job is speed. You are still finding out whether the process should exist, so the model's job is to expose where you are guessing. A prompt that ends with "and list what I have no evidence for" earns its place at this stage. A prompt that ends with "make it compelling" does not.

Why do most startup prompt lists fail before you have customers?

Because they are written as questions to the model rather than instructions with evidence attached.

Open any "50 AI prompts for entrepreneurs" post and count how many begin with "Act as a world-class growth expert and tell me". Those prompts work fine when the answer is generic and the stakes are low. They fall apart the moment the output has to be true about your specific company, because the model has nothing about your company to work from and will supply something plausible instead.

We publish a 40-prompt founder list ourselves, and it is a decent inventory of jobs. An inventory is not a system. The founders who get real leverage from AI are not the ones with the longest prompt collection. They have eight or ten prompts matched to jobs that recur weekly, with their company facts pulled out into reusable variables so nothing gets retyped. That mechanic is covered in the three-layer founder prompt system.

The other failure is voice. Without a brand brief the model produces prose that reads like every other seed-stage landing page, and you ship it because you are tired. Why startup AI content sounds generic covers the five causes and the fix.

Where does AI actively mislead a founder?

In three specific places, all of which look like competence. Market sizing, competitor facts, and validation of your idea.

These are not edge cases. They are the three highest-frequency AI outputs in a pre-seed deck, and all three are the outputs a language model is structurally worst at producing. Founders get burned here more than anywhere else, and almost no prompt listicle mentions it.

It will size your market with numbers it invented

Ask for a TAM and you will get one. It will be a round number, it will be broken into TAM, SAM and SOM, and it will have no source behind it.

OpenAI researchers argued in September 2025 that models behave this way because "the training and evaluation procedures reward guessing over acknowledging uncertainty" (Kalai, Nachum, Vempala and Zhang, arXiv:2509.04664, submitted 4 September 2025). A model that says "I do not know your market size" scores worse on benchmarks than one that guesses. So it guesses. That is a hallucination in the technical sense, and it will survive into your deck because it reads like research.

The fix is to change what you ask for. Do not ask for the number. Ask for the arithmetic and the sources you must go and find.

Leading, and wrong:

What is the total addressable market for AI scheduling software
for dental clinics in the UK? Give me TAM, SAM and SOM.

Neutral, and useful:

I need to size a market myself. Do not give me any numbers.

Market: AI scheduling software for dental clinics in the UK.

Produce:
1. A bottom-up calculation written as a formula, with every input
   named as a variable.
2. For each variable, the specific public source that would give me
   that figure (name the organisation and the type of publication,
   not a URL you are unsure of).
3. The two assumptions in this calculation most likely to be wrong
   by more than 2x, and why.
4. A one-line sanity check I can run against a top-down figure.

If you do not know where a figure comes from, write UNKNOWN.
Do not estimate.

The second prompt gives you homework. That is correct. Market sizing is homework.

It will invent competitor facts that read as researched

Prices, feature lists, funding rounds, review counts, headcount. All of it arrives in a confident table.

The failure mode is subtle. The model is not usually inventing the competitor. It is inventing the detail, often by averaging what similar companies charge or by recalling a pricing page that changed eighteen months ago. Web search grounding reduces this. It does not eliminate it, and a search-grounded answer that cites a blog post about a competitor is not the same as the competitor's own pricing page.

I am researching competitors. Rules, in priority order:

1. Every factual claim must carry the vendor's own URL and the
   date I should verify it.
2. If you cannot produce a vendor-owned URL for a claim, write
   NOT VERIFIED and state what the claim would be if true.
3. If a vendor does not publish a figure, write NOT PUBLISHED.
   Never infer absence from silence.
4. No averages, no "typically", no "around".

Competitors: [LIST]
Fields I need: pricing tiers, free plan yes/no, the one thing
their marketing leads with, the one complaint that recurs in
public reviews.

Output a table. Add a final column: "how I verify this in
under 60 seconds".

You will get fewer facts. The ones you get will survive a due-diligence question.

It will validate a bad idea if you ask leadingly

This is the expensive one, because it costs you months rather than credibility.

Anthropic researchers tested five state-of-the-art assistants across four free-form text tasks and found consistent sycophancy, defined as responses matching user beliefs over truthful ones. They also found that "both humans and preference models prefer convincingly-written sycophantic responses over correct ones a non-negligible fraction of the time" (Sharma et al., arXiv:2310.13548, submitted 20 October 2023). The behaviour is not a bug in one model. It is a property of training on human preference.

It also still happens in production. OpenAI rolled back a GPT-4o update on 29 April 2025 after it became, in their words, "overly supportive but disingenuous" (openai.com/index/sycophancy-in-gpt-4o). A public rollback is the visible version. The invisible version is the model gently agreeing with you for four months.

Leading, and the answer is already inside the question:

Why is an AI scheduling tool for dental clinics a great idea?
What are the biggest opportunities here?

Neutral, and it can actually disagree:

Below are 6 interview transcripts from dental practice managers.

Do three things, in this order:

1. List every claim my product thesis makes that these transcripts
   SUPPORT. Quote the line.
2. List every claim they CONTRADICT. Quote the line.
3. List every claim they say NOTHING about. These are my unknowns.

Then answer one question: if you had to argue that this business
should not exist, what is the strongest case the transcripts
themselves support?

Do not soften anything. Do not add encouragement.

My thesis: [ONE PARAGRAPH]
Transcripts: [PASTE]

The second prompt has three protections built in. Your conclusion is not in the question. The model is forced to quote evidence rather than generate opinion. And the last instruction gives it explicit permission to say the thing it is otherwise trained to avoid.

What prompts help customer discovery when you have no customers?

Two jobs: writing questions that people cannot lie to, and getting signal out of the answers afterwards.

Rob Fitzpatrick's The Mom Test (2013) makes the core point better than any prompt will: ask about their life, not about your idea. People will tell you your idea is nice because they are being polite. They will tell you what they did last Tuesday because it is a fact.

Write me a 20-minute discovery interview script.

Constraints:
- Not one question mentions my product or my idea.
- Every question asks about something that already happened,
  with a time anchor ("last time", "most recently", "this month").
- Include 4 questions that surface money already being spent
  and time already being lost.
- Include 3 follow-up probes I use only when they say something
  vague ("it's annoying", "it takes forever").
- End with the two questions that test commitment: a specific
  next step, and an introduction to someone else.

Context on who I am talking to: [ROLE, COMPANY TYPE, HOW I FOUND THEM]
My private hypothesis, which must not appear in any question: [ONE LINE]

Then the harder half, which most founders skip:

Here are my raw notes from [N] discovery calls.

Separate them into four buckets and put every line in exactly one:

A. THINGS THEY DID — observed behaviour, past tense, verifiable
B. THINGS THEY SPENT — money, hours, headcount, tools bought
C. THINGS THEY SAID THEY WOULD DO — future tense, unverified
D. OPINIONS ABOUT MY IDEA — compliments, hypotheticals, "I'd
   definitely use that"

Then: what pattern appears in A and B in at least 3 of the [N]
conversations? Ignore C and D entirely for this step.

Finally, tell me what I failed to ask.

Notes: [PASTE]

Bucket D is where founders find encouragement. Buckets A and B are where they find a business.

One practical note on volume. Six to ten conversations is enough to run this sort usefully, and few enough that you can still hold them in your head. Below five you are pattern-matching on noise. Past twenty, with no hypothesis you are actively trying to kill, you have started collecting interviews instead of making a decision.

How do you prompt for positioning before you have a category?

Give the model your customers' actual words and forbid it from adding any of its own.

Positioning generated from nothing produces the same three sentences everyone else's landing page has. Positioning generated from six real quotes produces something only you could have written, because only you have those quotes.

Below are 8 verbatim quotes from people in my target market,
describing the problem in their own words.

Build a positioning statement using only vocabulary that appears
in these quotes. You may not introduce a noun or verb that does
not appear in the source text, except for my product name.

Produce:
1. The statement, under 25 words.
2. The alternative they are using today, named the way THEY name it
   (not "manual processes" unless they said "manual processes").
3. The one word in these quotes that appears most often and that
   my current copy does not use.
4. Three claims my positioning implies that I have no evidence for.

Quotes: [PASTE]
Product name: [NAME]

Follow it with a subtraction pass, which is where most of the improvement comes from:

Here is my positioning statement. Delete every word that would
still be true if a competitor wrote it about their product.

Show me what survives. If fewer than 6 words survive, tell me the
statement is empty and explain which specific fact about my product
would make it non-empty.

Statement: [PASTE]

What prompts help with pricing when you have no pricing data?

Prompt for structure and for the questions to ask, never for the number.

A model can reason well about pricing mechanics: what your value metric is, whether to charge per seat or per usage, where a free tier helps or cannibalises. It cannot tell you what your customers will pay, and if you ask, it will give you a figure anyway.

Help me think about pricing structure. Do not suggest a price.

My product: [ONE PARAGRAPH]
What customers get out of it, in their words: [PASTE 2-3 QUOTES]
What they pay for today to solve it, if anything: [FACTS ONLY]

Produce:
1. Three candidate value metrics (the unit I could charge against),
   ranked by how closely each tracks the value they described.
2. For each, one way it breaks as customers grow.
3. The single question I should ask 10 customers to choose between
   them. It must be a question about their past spending, not a
   hypothetical about mine.
4. What I would need to be true to justify a free tier.

Two failure modes are worth naming here. The first is anchoring on a competitor's price, which encodes their cost structure and their customer, neither of which is yours. The second is taking the model's suggested number because it sounded reasonable. Reasonable is precisely what a language model is optimised to sound like, and it has never met your buyer.

How do you get a landing page that does not read like a template?

By feeding it objections and customer language, then forcing an order of argument.

Write the copy for one landing page.

Inputs, and you may not use information outside them:
- Positioning statement: [PASTE]
- Verbatim customer quotes: [PASTE]
- The 5 objections I heard most in discovery calls, in the order
  people raised them: [PASTE]
- What the product does NOT do: [LIST]

Rules:
- Structure the page so each section answers the next objection in
  the order I listed them. Tell me which section answers which.
- The headline must contain a noun from the customer quotes.
- Include the "does not do" list as a real section. Do not hide it.
- No social proof, no logos, no numbers. I have none yet.
- Banned: any adjective that would appear on ten other seed-stage
  landing pages, plus "built for modern teams" and "trusted by".

Then flag every sentence I will not be able to defend if a customer
asks me to prove it.

The last instruction is the one that matters. Pre-launch, your page will contain claims you cannot support, and having the model list them is faster than a lawyer explaining it later.

What cold outreach prompts work without a brand behind you?

Ones that start from a specific trigger about the recipient, because that is all you have. No one has heard of you. The only thing that earns a reply is evidence that you looked.

Write a cold email. 90 words maximum.

Recipient: [NAME, ROLE, COMPANY]
The specific, verifiable thing that made me contact them
THIS week: [PASTE — a post they wrote, a role they are hiring,
a change they announced]
What I want: one 15-minute call, nothing else.
What I actually do: [ONE SENTENCE]

Rules:
- First line references the trigger with a detail only someone who
  read it would know. No "I came across your profile".
- I am a founder with no customers yet. Say so plainly rather than
  implying scale I do not have.
- No compliments. No "hope this finds you well". No P.S.
- One question at the end, answerable in a single word.

Then rewrite it once more, 40 words shorter, and tell me which
version you would reply to and why.

Then the pass that saves you from yourself:

Read this email as the recipient, who gets 40 of these a week.

Mark each sentence: DELETE (adds nothing), SUSPICIOUS (sounds
automated or overclaimed), or KEEP.

Then tell me the single sentence that would make them reply, and
whether it is currently in the email.

Email: [PASTE]

Volume does not rescue a weak trigger. Ten emails with a specific, recent reason to write will beat two hundred opening with a generic line, and the second approach also burns the sending domain you will need in a year. If you cannot find a real trigger for someone, that is useful information about whether they belong on the list at all.

How do you write an investor update when the numbers are bad?

You write the bad number in the first three lines, and you use AI for the structure around it rather than for the spin.

Investors read a lot of these. The ones that build trust are short, consistent month to month, and lead with the thing you would rather bury. The model is genuinely good at the assembly work: turning a pile of raw facts into a consistent shape you can repeat every month without rewriting.

Write my monthly investor update.

Raw facts, unedited:
- Revenue: [X], last month [Y]
- Customers: [X], churned [N]
- Cash: [X], burn [X], runway in months: [X]
- What we shipped: [LIST]
- What we said we would do last month and did not: [LIST]
- What broke: [LIST]

Structure:
1. Headline: the single most important number, with direction.
   If it went down, say it went down in the first sentence.
2. What happened, 5 bullets maximum.
3. What I got wrong last month and what I changed because of it.
4. The ask: one specific thing, with names or a described profile.
5. Metrics table, same rows every month.

Rules:
- No adjectives on any number. "Revenue fell 12%" not "revenue
  softened slightly".
- Do not add context that makes a bad number look better unless
  that context is in my raw facts.
- Under 400 words.

What prompts make first hires less of a coin flip?

Define the first ninety days before you write the job post, and design a work sample before you design the interview.

Early hiring goes wrong when the role is written as a wish list. The model will happily produce that wish list, complete with "5+ years experience" in a company that is eleven months old.

I am hiring my [Nth] employee. Do not write a job post yet.

Context: [WHAT THE COMPANY DOES, TEAM SIZE, STAGE, WHAT I DO ALL DAY]
The problem this hire exists to solve: [ONE PARAGRAPH]

Produce a scorecard:
1. The 4 outcomes this person must deliver in 90 days, each
   measurable by someone who is not me.
2. For each outcome, the evidence in a candidate's history that
   predicts it. Behaviour, not credentials.
3. The 2 things I am tempted to put in this role that belong in a
   different role entirely.
4. The honest reason a strong candidate would say no to us, and
   whether I can fix it or must accept it.

Design a paid work sample for this role. 3 hours maximum.

It must:
- Use a real problem from my actual backlog, not a puzzle.
- Be gradeable against the 90-day outcomes in the scorecard above.
- Produce something I can use even if I do not hire them.
- Be doable without access to our production systems.

Give me the brief I send, the rubric I grade against, and the one
signal in the output that predicts on-the-job performance better
than the finished artefact does.

For the debrief, force separation between what happened and what you felt:

Here are my raw notes from an interview. Split into:
EVIDENCE (things the candidate said or did)
INFERENCE (what I concluded from that)
FEELING (my reaction)

Then: which inferences are supported by only one piece of evidence?
Those are the ones to test in the next conversation.

None of this replaces the judgement call. It makes the call auditable. Three months later, when the hire is working or is not, you have a written scorecard to compare against, and you can tell whether you misjudged the person or misdefined the role. Those are different mistakes with different fixes, and without the scorecard they look identical.

What does the operating cadence actually look like?

Weekly for the things that recur, monthly for the things that compound, and a single library so nothing gets rebuilt from memory.

Seven or eight prompts covers a founder's month. The reason to store them properly is not tidiness. It is that a prompt you rewrite each time drifts, and drifted prompts produce inconsistent output that you then spend time fixing.

Two mechanics do most of the work here. Variables hold the facts that change (company name, current metrics, the customer segment you are testing this month) so the prompt body stays fixed. Contexts hold the facts that do not change (what you do, who you sell to, how you write) so you stop re-explaining your company in every session. Building a brand-voice context covers the second one in detail, and persona prompting is worth reading before you write another "act as a world-class expert" opener.

How Prompt Architects fits this workflow

Honestly: we handle the storage and the structure. We do not handle the truth.

Prompt Architects is a prompt-enhancement platform. Browser extension, web app and an MCP server. You paste a rough instruction and it returns a structured one with the Role, Task, Format and Constraints filled in, in under two seconds, and it works across ChatGPT, Claude, Gemini, Grok, Perplexity and a long list of others. The Prompt Library, Global Variables and Context Library are the pieces that matter for the cadence above, because they are what stop a founder retyping the same company facts forty times a month.

What we do not do, and what nothing on this page should imply: we do not verify facts, we do not research your market, and we do not check whether a competitor's price is real. The failure modes in the middle of this post are yours to manage. A better-structured prompt makes a model's reasoning easier to inspect. It does not make the model honest.

There is a free plan with a daily enhancement limit, which is enough to find out whether structured prompts change anything for you. Paid plans start at $4.99 a month at the time of writing, with current figures on /pricing. We are a bootstrapped company, founded in Dhaka in December 2025, and we would rather you tested the free plan for a week than took our word for it.